Replies to post #10481 on Intel Corp (INTC)

[Jerry R]

SmallPops - That data shows Intel had a better improvement than IBM did going to 90nm.

This, of course, pushes the inquiry back one step to ask: what is the purpose of the extra 37M non-cache transitors when going from Northwood to Prescott?

It does seem to confirm what SemiconEng has been saying - there is nothing wrong with Intel's 90nm process.

[wbmw]

Smallpops, Re: That data shows Intel had a better improvemnet than IBM did going to 90nm.

I never said the process was at fault. I blame the micro-architecture. Intel decided to implement a 31-stage pipeline when they should have added all of Prescott's features to a Northwood sized 20-stage pipeline. It would have given them much greater IPC, and based on the data you posted, it would have likely given them lower power. I bet it would clock higher, too. It would have been the ideal solution; instead, the Prescott architects, probably led by marketing, decided to push for gigahertz at a time when the thermal limit has finally hit a wall. The end result is simply disastrous - a technically inferior solution. I would like to see Intel remedy this soon, either by going back to their 20-stage solution, or better yet, a Pentium M derived solution.

[chipguy]

Are you sure that they really messed it up

Your data shows it. Why burn more than 2x more logic
transistors to barely recover IPC lost in stretching the
pipeline by nearly 50% when the power *increase* per
GHz despite the shrink (!!!) effectively disposes of
the clock scalability benefit from that longer pipeline?

Look at the last shrink - a 2 GHz Northwood had a TDP
of 52 W compared to 92 W for a 2 GHz Willamette AND
the Northwood performed roughly 10% better per clock.
A simple shrink of Northwood to 90 nm with a 1 MB L2
would have provided 1) lower power per GHz, 2) clearly
better performance per GHz, and 3) much higher clock
rate scalability (although the gain would be less than
Northwood vs Willamette for any number of reasons).
The Prescott so far fails to achieve *any* of those 3
criteria - an incredible "achievement".